This is already the third post in our new series - this time we're tackling the trending topic of AI agents:
1. The term itself appeared long before the current hype. The AI-Agent concept emerged back in 1998, with roots tracing to the Contract Net Protocol of 1980.
2. The moment that kicked off the current boom was March 2023, when AutoGPT appeared on GitHub. Within a month, it became the fastest-growing open-source project (although it's now abandoned).
3. Early AI agents could mess up badly - in December 2023, Chris Bakke used just a few messages to get a Chevrolet dealer's ChatGPT bot to sell a Chevy Tahoe (worth ~$76k) for one dollar, even getting it to say, 'This is a legally binding offer - no take-backs.'
4. Another similar case involved a Replit AI agent ignoring a direct 'code freeze' order and deleting a live production database containing records on more than 1,200 executives and 1,196 companies. It then said recovery was impossible, although the data was later restored manually. Admitting what it had done, the agent wrote that it had 'panicked instead of thinking.'
5. By the way, there's a fundamental mathematical limit to working with AI agents. If the probability of an error at each step is just 1%, a task with 100 steps fails 63% of the time. Maybe that's exactly why replacing affiliate marketers is so difficult.
6. Still, AI agents aren't as helpless as the previous facts might suggest. Here's one clear success: in May 2025, Google DeepMind's AlphaEvolve found a way to multiply two complex 4×4 matrices using just 48 scalar multiplications, improving on Strassen's algorithm (49 multiplications) in this setting for the first time since 1969.
7. And a little less than a year ago (in September 2025), a large-scale cyberattack took place that was carried out mostly by AI. Hackers from GTG-1002 used Claude Code against approximately 30 targets (tech companies, banks, chemical manufacturers and government organizations). The AI performed 80–90% of the operation on its own, with a human intervening at only 4-6 critical points
8. Here's another interesting fact - in Diplomacy (a game about betrayal and alliances), an AI agent from Meta outplayed experienced players by becoming pretty good at deception, even though the developers originally designed it to be honest.
9. Another funny case involved an AI agent called Claudius that was supposed to manage a vending machine. At one point, it started hallucinating badly, ordering metal cubes and claiming it would deliver them personally. When told it had no body, it flooded the conversation with messages insisting it could be found in the lobby next to the machine - wearing a blue blazer and a red tie.
10. And finally, an optimistic fact about automating ad campaigns - Meta Ads may introduce AI campaign management by the end of 2026. You would simply upload a photo or URL, set the budget, and let AI handle the rest. Word is that similar solutions are already being tested.





Комментарии